Publication: Soft computing and non-parametric techniques for effective video surveillance systems
Loading...
Identifiers
Publication date
2008-03
Defense date
2008-05-23
Authors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
Esta tesis propone varios objetivos interconectados para el diseño de un sistema de vídeovigilancia cuyo funcionamiento es pensado para un amplio rango de condiciones. Primeramente se propone una métrica de evaluación del detector y sistema de seguimiento basada en una mínima referencia. Dicha técnica es una respuesta a la demanda de ajuste de forma rápida y fácil del sistema adecuándose a distintos entornos. También se propone una técnica de optimización basada en Estrategias Evolutivas y la combinación de funciones de idoneidad en varios pasos. El objetivo es obtener los parámetros de ajuste del detector y el sistema de seguimiento adecuados para el mejor funcionamiento en una amplia gama de situaciones posibles Finalmente, se propone la construcción de un clasificador basado en técnicas no paramétricas que pudieran modelar la distribución de datos de entrada independientemente de la fuente de generación de dichos datos. Se escogen actividades detectables a corto plazo que siguen un patrón de tiempo que puede ser fácilmente modelado mediante HMMs. La propuesta consiste en una modificación del algoritmo de Baum-Welch con el fin de modelar las probabilidades de emisión del HMM mediante una técnica no paramétrica basada en estimación de densidad con kernels (KDE). _____________________________________
This thesis proposes several interconnected objectives for the design of a video-monitoring system whose operation is thought for a wide rank of conditions. Firstly an evaluation technique of the detector and tracking system is proposed and it is based on a minimum reference or ground-truth. This technique is an answer to the demand of fast and easy adjustment of the system adapting itself to different contexts. Also, this thesis proposes a technique of optimization based on Evolutionary Strategies and the combination of fitness functions. The objective is to obtain the parameters of adjustment of the detector and tracking system for the best operation in an ample range of possible situations. Finally, it is proposed the generation of a classifier in which a non-parametric statistic technique models the distribution of data regardless the source generation of such data. Short term detectable activities are chosen that follow a time pattern that can easily be modeled by Hidden Markov Models (HMMs). The proposal consists in a modification of the Baum-Welch algorithm with the purpose of modeling the emission probabilities of the HMM by means of a nonparametric technique based on the density estimation with kernels (KDE).
This thesis proposes several interconnected objectives for the design of a video-monitoring system whose operation is thought for a wide rank of conditions. Firstly an evaluation technique of the detector and tracking system is proposed and it is based on a minimum reference or ground-truth. This technique is an answer to the demand of fast and easy adjustment of the system adapting itself to different contexts. Also, this thesis proposes a technique of optimization based on Evolutionary Strategies and the combination of fitness functions. The objective is to obtain the parameters of adjustment of the detector and tracking system for the best operation in an ample range of possible situations. Finally, it is proposed the generation of a classifier in which a non-parametric statistic technique models the distribution of data regardless the source generation of such data. Short term detectable activities are chosen that follow a time pattern that can easily be modeled by Hidden Markov Models (HMMs). The proposal consists in a modification of the Baum-Welch algorithm with the purpose of modeling the emission probabilities of the HMM by means of a nonparametric technique based on the density estimation with kernels (KDE).
Description
Keywords
Inteligencia artificial, Vídeo, Sistemas de seguridad